TL;DR: OpenCode is the free, open-source alternative to Claude Code, and by default it can run shell commands directly on your machine. I kept thinking about the same failure modes: a misread symlink, an rm -rf pointed at the wrong path, or a postinstall script quietly modifying files I never intended to touch. The solution wasn't watching the agent more closely—it was moving its shell into a disposable Tensorlake sandbox while keeping the rest of my workflow local. I cover the full setup, configuration, and why this changed how I use coding agents in the Medium article linked below.

You type what you want. OpenCode reads your codebase, edits files, and runs shell commands on its own.

Same idea as Claude Code. Free, open source, not locked to one model.

Great, until you sit with what "runs shell commands on its own" actually means.